EnergyLM-SFT

EnergyLM 3B fine-tuned with supervised fine-tuning (SFT) on identity and energy-domain data.

Model

• Architecture: EnergyLM (3B params, 36 layers, 2048 hidden, GQA)
• Base: EOS-Base pretrained on FineWeb-edu
• Fine-tuning: SFT on Safire/EnergyLM identity and conversational data

Usage

HuggingFace (Transformers)

Install the EnergyLM package and load:

# Requires: pip install transformers
# Clone this repo for model registration
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("ksjpswaroop/energylm-sft")
tokenizer = AutoTokenizer.from_pretrained("ksjpswaroop/energylm-sft")

llama.cpp / GGUF

Download the GGUF file (if available) and run with llama.cpp or llama-cpp-python:

pip install llama-cpp-python
python -c "
from llama_cpp import Llama
llm = Llama(model_path='energylm-sft.gguf', n_ctx=4096)
prompt = '<|user|>\nWho are you?\n<|assistant|>\n'
out = llm(prompt, max_tokens=256, stop=['<|end_of_text|>', '<|user|>'])
print(out['choices'][0]['text'])
"

Identity

This model identifies as EOS-chat / EnergyLM, developed by Safire's AI team, with a focus on energy sector applications.

License

Apache 2.0